Serena

Serena is a semantic code navigation MCP that plugs language servers into your AI agent. Instead of your LLM doing dumb string searches, it gets proper symbol lookups, cross-file references, and rename refactors across 40-plus languages. It's the difference between asking Claude to "find where this function is called" and getting grep output versus getting an actual call hierarchy. I've used it most heavily in TypeScript and Python repos where cross-file imports get messy. The killer feature is reliable rename: tell it to rename a method and it propagates correctly through the entire codebase, respecting scope and imports. That's not trivial. Most LLM-driven refactors either miss references or rename too much. Serena leans on the same language servers your editor uses, so it gets the semantics right. The workflow I reach for: before any large refactor, I run a usage lookup on the symbol I'm about to change. Serena returns every reference with file paths and line numbers. I can see the blast radius before I commit. Then I use the rename tool if it's straightforward, or at least I know what to review manually. Quirks: it's a Python package installed via uvx, so you need Python 3.10 or later. The language server backends have to be installed separately for each language you care about (pyright for Python, typescript-language-server for TS, etc.). The README lists them, but it's not automatic. If you're working in a language without a mature LSP, you're back to grep. Who shouldn't bother: if you only write single-file scripts or work in a language without LSP support, this is overkill. Also, if you're already happy with your editor's refactor tools and don't need an agent to do semantic ops, skip it. But if you're using Claude Desktop or Cursor to navigate a real codebase and you're tired of it hallucinating references, Serena is the fix. It's not flashy, but it's the kind of tool that quietly makes your agent competent at the one thing it should be good at: understanding code structure.